Summary
OpenAI says CodeRabbit uses GPT-6.1 Sol as the primary model in its security agent, applying it to pull-request reviews and broader repository scans. The workflow can also attempt sandboxed exploit checks and draft fixes for human review.
OpenAI said on October 26, 2026, that CodeRabbit is using GPT-6.1 Sol as the primary model in its security agent. The system is intended to examine risks that can span files, components and dependencies, rather than focusing only on the lines changed in a pull request. OpenAI says CodeRabbit combines Sol with other frontier models and security tools.
Extending review beyond changed lines
A pull request is a proposed set of code changes submitted for review. Those changes can affect shared libraries, dependencies or safeguards elsewhere in an application, so understanding their security impact may require looking beyond the edited files.
OpenAI says Sol helps CodeRabbit connect behaviour across files and components to identify potential security issues. In the pull-request workflow, it analyses changes as developers work. CodeRabbit’s Deep Scan broadens the investigation to repositories, dependencies and interactions between components. CodeRabbit Secure can extend that analysis to linked repositories, such as a shared authentication library, to help teams assess how a weakness could affect other parts of a system.
CodeRabbit reports that it has used Sol to flag security-related issues across thousands of repositories and organisations. OpenAI describes the review as independent of the agent that wrote the code, and as a complement to established scanners.
From findings to proposed fixes
CodeRabbit Secure can attempt to verify user-selected findings by testing for an exploit in an isolated sandbox, rather than against production systems. For supported issues, it can also draft a pull request proposing a fix. A person reviews and approves the proposed change; human reviewers remain responsible for deciding what is shipped.
The described workflow links broader code-context analysis with optional verification and a human-reviewed remediation path. OpenAI’s announcement does not report measured detection or false-positive rates, so it describes how the tools are intended to work rather than quantified security performance.